/*
 * Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 *
 * SPDX-License-Identifier: MIT
 */

#include <hip_test_common.hh>
#include <resource_guards.hh>

#include "__syncthreads_count_negative_kernels_rtc.hh"
#include "syncthreads_common.hh"

/**
 * @addtogroup __syncthreads_count __syncthreads_count
 * @{
 * @ingroup SyncthreadsTest
 */

/**
 * Test Description
 * ------------------------
 *    - Basic synchronization test for `__syncthreads_count`.
 *
 * Test source
 * ------------------------
 *    - unit/syncthreads/__syncthreads_count.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
HIP_TEST_CASE(Unit___syncthreads_count_Positive_Basic) {
  const auto kGridSize = 2;
  const auto kBlockSize = GENERATE(13, 32, 64, 513);

  LinearAllocGuard<int> out_alloc(LinearAllocs::hipMallocManaged, sizeof(int) * kGridSize);

  HipTest::launchKernel(SyncthreadsKernel<SyncthreadsKind::kCount>, kGridSize, kBlockSize,
                        sizeof(int) * kBlockSize, nullptr, out_alloc.ptr());
  HIP_CHECK(hipDeviceSynchronize());

  for (int i = 0; i < kGridSize; ++i) {
    REQUIRE(out_alloc.host_ptr()[i] == kBlockSize * (kBlockSize + 1) / 2);
  }
}

/**
 * Test Description
 * ------------------------
 *    - Test `__syncthreads_count` with 0 as the predicate for all threads.
 *
 * Test source
 * ------------------------
 *    - unit/syncthreads/__syncthreads_count.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
HIP_TEST_CASE(Unit___syncthreads_count_Positive_Predicate_Zero) {
  const auto kGridSize = 2;
  const auto kBlockSize = GENERATE(13, 32, 64, 513);

  LinearAllocGuard<int> out_alloc(LinearAllocs::hipMallocManaged,
                                  sizeof(int) * kGridSize * kBlockSize);

  HipTest::launchKernel(SyncthreadsZeroKernel<SyncthreadsKind::kCount>, kGridSize, kBlockSize, 0,
                        nullptr, out_alloc.ptr());
  HIP_CHECK(hipDeviceSynchronize());

  for (int i = 0; i < kGridSize * kBlockSize; ++i) {
    REQUIRE(out_alloc.host_ptr()[i] == 0);
  }
}

/**
 * Test Description
 * ------------------------
 *    - Test `__syncthreads_count` with 1 as the predicate for all threads.
 *
 * Test source
 * ------------------------
 *    - unit/syncthreads/__syncthreads_count.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
HIP_TEST_CASE(Unit___syncthreads_count_Positive_Predicate_One) {
  const auto kGridSize = 2;
  const auto kBlockSize = GENERATE(13, 32, 64, 513);

  LinearAllocGuard<int> out_alloc(LinearAllocs::hipMallocManaged,
                                  sizeof(int) * kGridSize * kBlockSize);

  HipTest::launchKernel(SyncthreadsOneKernel<SyncthreadsKind::kCount>, kGridSize, kBlockSize, 0,
                        nullptr, out_alloc.ptr());
  HIP_CHECK(hipDeviceSynchronize());

  for (int i = 0; i < kGridSize * kBlockSize; ++i) {
    REQUIRE(out_alloc.host_ptr()[i] == kBlockSize);
  }
}

/**
 * Test Description
 * ------------------------
 *    - Test `__syncthreads_count` with 0 as the predicate for even threads, and 1 as the predicate
 * for odd threads.
 *
 * Test source
 * ------------------------
 *    - unit/syncthreads/__syncthreads_count.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
HIP_TEST_CASE(Unit___syncthreads_count_Positive_Predicate_OddEven) {
  const auto kGridSize = 2;
  const auto kBlockSize = GENERATE(13, 32, 64, 513);

  LinearAllocGuard<int> out_alloc(LinearAllocs::hipMallocManaged,
                                  sizeof(int) * kGridSize * kBlockSize);

  HipTest::launchKernel(SyncthreadsOddEvenKernel<SyncthreadsKind::kCount>, kGridSize, kBlockSize, 0,
                        nullptr, out_alloc.ptr());
  HIP_CHECK(hipDeviceSynchronize());

  for (int i = 0; i < kGridSize * kBlockSize; ++i) {
    REQUIRE(out_alloc.host_ptr()[i] == kBlockSize / 2);
  }
}

/**
 * Test Description
 * ------------------------
 *    - Test `__syncthreads_count` with a negative predicate.
 *
 * Test source
 * ------------------------
 *    - unit/syncthreads/__syncthreads_count.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
HIP_TEST_CASE(Unit___syncthreads_count_Positive_Predicate_Negative) {
  const auto kGridSize = 2;
  const auto kBlockSize = GENERATE(13, 32, 64, 513);

  LinearAllocGuard<int> out_alloc(LinearAllocs::hipMallocManaged,
                                  sizeof(int) * kGridSize * kBlockSize);

  HipTest::launchKernel(SyncthreadsNegativeKernel<SyncthreadsKind::kCount>, kGridSize, kBlockSize,
                        0, nullptr, out_alloc.ptr());
  HIP_CHECK(hipDeviceSynchronize());

  for (int i = 0; i < kGridSize * kBlockSize; ++i) {
    REQUIRE(out_alloc.host_ptr()[i] == kBlockSize);
  }
}

/**
 * Test Description
 * ------------------------
 *    - Test `__syncthreads_count` with the thread ID as the predicate.
 *
 * Test source
 * ------------------------
 *    - unit/syncthreads/__syncthreads_count.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
HIP_TEST_CASE(Unit___syncthreads_count_Positive_Predicate_Id) {
  const auto kGridSize = 2;
  const auto kBlockSize = GENERATE(13, 32, 64, 513);

  LinearAllocGuard<int> out_alloc(LinearAllocs::hipMallocManaged,
                                  sizeof(int) * kGridSize * kBlockSize);

  HipTest::launchKernel(SyncthreadsIdKernel<SyncthreadsKind::kCount>, kGridSize, kBlockSize, 0,
                        nullptr, out_alloc.ptr());
  HIP_CHECK(hipDeviceSynchronize());

  for (int i = 0; i < kGridSize * kBlockSize; ++i) {
    REQUIRE(out_alloc.host_ptr()[i] == kBlockSize - 1);
  }
}

/**
 * Test Description
 * ------------------------
 *    - Real-time compiles kernels that pass invalid arguments to `__syncthreads_count`.
 *
 * Test source
 * ------------------------
 *    - unit/syncthreads/__syncthreads_count.cc
 * Test requirements
 * ------------------------
 *    - HIP_VERSION >= 5.2
 */
HIP_TEST_CASE(Unit___syncthreads_count_Negative_Parameters_RTC) {
  hiprtcProgram program{};

  HIPRTC_CHECK(hiprtcCreateProgram(&program, kSyncthreadsCountSource,
                                   "__syncthreads_count_negative.cc", 0, nullptr, nullptr));
  hiprtcResult result{hiprtcCompileProgram(program, 0, nullptr)};

  // Get the compile log and count compiler error messages
  size_t log_size{};
  HIPRTC_CHECK(hiprtcGetProgramLogSize(program, &log_size));
  std::string log(log_size, ' ');
  HIPRTC_CHECK(hiprtcGetProgramLog(program, log.data()));
  int error_count{0};

  int expected_error_count{2};
  std::string error_message{"error:"};

  size_t n_pos = log.find(error_message, 0);
  while (n_pos != std::string::npos) {
    ++error_count;
    n_pos = log.find(error_message, n_pos + 1);
  }

  HIPRTC_CHECK(hiprtcDestroyProgram(&program));
  HIPRTC_CHECK_ERROR(result, HIPRTC_ERROR_COMPILATION);
  REQUIRE(error_count == expected_error_count);
}

/**
 * End doxygen group SyncthreadsTest.
 * @}
 */